Defective unmanned aerial vehicles (UAVs), like other malfunctioning machinery, often emit abnormal sounds. Early detection of such acoustic abnormalities enables timely countermeasures to prevent equipment loss and human casualties. This paper proposes a deep learning scheme to passively detect abnormal sounds emitted from UAVs, using only a single ground-based acoustic vector sensor (AVS). This is because lone AVS measured sound intensity spectrogram can already provide information on both the sound source’s polar-azimuthal direction and abnormality time-frequency signature. This paper integrates a "convolutional autoencoder" with an "attention module", for UAV abnormality detection. Field experiments, with a UAV in flight and a single AVS on the ground, validate the effectiveness of the proposed method, achieving an F1-score (a machine-learning accuracy metric) exceeding 95% for successful anomaly detection.